Human approval for AI agents: what to approve, what to let run
AI agents that pay, post or delete need a human yes. Which actions need approval, how standing permissions work, and why you need a kill switch.
The fastest way to lose trust in AI agents is one bad action: an email to the wrong list, a deleted folder, an ad budget raised by mistake. The fix isn't to stop using agents. It's to put a person in the loop at the right moments, and nowhere else.
The rule of thumb
Let the AI do anything that is easy to undo or only reads. Ask a person before anything that is hard to undo or reaches other people.
| Usually fine without asking | Ask first |
|---|---|
| Reading dashboards and reports | Paying or buying anything |
| Drafting emails, posts and replies | Sending messages, emails or posts to people |
| Updating your own notes and trackers | Deleting files, posts or contacts |
| Downloading reports | Changing passwords, settings or permissions |
| Searching the web | Using your bank or payment logins |
AgentDaftar ships with every item in the right column set to ask. You can relax them per office once you trust a workflow.
Make approval fast
An approval nobody sees is just a stuck task. Approvals work when they:
- Reach your phone. AgentDaftar sends them to Telegram with buttons: Yes, once, Always for this task, No.
- Say exactly what will happen. The kind of action, the task it belongs to, the recipients and the amount.
- Expire. Unanswered requests expire after 24 hours, so nothing old gets approved by accident.
Standing permissions, done safely
Approving the same daily report every morning gets old. A standing permission says "yes to this, from now on". The danger is a permission that's too broad: a misled AI could stretch "send the daily report" into "send anything to anyone".
So keep standing permissions narrow. In AgentDaftar, tapping Always for this task creates one that covers:
- the same kind of action,
- for the same task,
- to the same recipients,
- at most so many times a day and up to an amount,
- until an end date (90 days by default).
Anything outside those limits is asked again.
Every team needs a kill switch
When something looks wrong, you don't want to find the right setting. You want one button. AgentDaftar's kill switch stops every AI at once: their actions are refused and running commands end. It's in the panel, and on Telegram as /stop (or /stop pc1 for one employee). /resume starts them again.
Keep a record
Approvals are half the story; the other half is being able to check what happened. Keep an activity log per agent, screen recordings of AI sessions, and a history of who changed which rule. When something goes wrong, you'll know exactly what and why.
Start strict, then relax
Begin with everything risky on ask. Watch a workflow for a week. When you've approved the same thing ten times without changes, give it a narrow standing permission. That's how teams go from nervous to confident with AI agents without a single bad surprise.
See how it works in AgentDaftar approvals, or read what an AI employee is.